Exploring Value of Synthetic Data from Overhead Perspective
summary
RarePlanes is an open dataset combining 14,700 hand-annotated aircraft from 253 Maxar WorldView-3 scenes with 50,000 synthetic images and ~630,000 aircraft annotations generated by AI.Reverie, each aircraft labeled with 10 fine-grain attributes (length, wingspan, wing shape/position, propulsion, engines, stabilizers, canards, role). Experiments show transfer-learning from synthetic then real cuts required real labels dramatically, boosting rare-object detection F1 from ~0.05 to ~0.35 for classes with ≤10 examples and near-human recall (~98%) on common classes. Data and code are on AWS S3 and GitHub.
AI-written summary · about
chapters (11)
- 0:00Rare Planes project intro
- 1:13Dataset overview and purpose
- 3:00Data collection process
- 5:04Dataset attributes and diversity
- 7:06Label taxonomy and aircraft
- 9:12Detection tasks and results
- 11:20Synthetic data augmentation
- 13:06Rare object detection challenges
- 14:50Fine-tuning for rare objects
- 16:36Dataset access resources
- 18:00Q&A
related talks
20mSynthetic Data: The new Super Power for AI
30m🤯 No data? No problem! Synthetic data to the rescue
33mNext-generation tools for synthetic data and AI training
58mUsing Unity to Generate Synthetic data and Accelerate Computer Vision Training
01h 00mThe Synthetic Data Advantage: How It Helps You to Get Both AI Privacy & Explainability
45mSynthetic Data for Machine Learning Applications
29mSynthetic Data in Vision AI: Theory and Practice
28mData Curation: Transforming Bytes into AI Gold